Leukocyte Image Segmentation Using Novel Saliency Detection Based on Positive Feedback of Visual Perception

This paper presents a novel method for salient object detection in nature image by simulating microsaccades in fixational eye movements. Due to a nucleated cell usually stained that is salient obviously, the proposed method is suitable to segment nucleated cell. Firstly, the existing fixation prediction method is utilized to produce an initial fixation area. Followed EPELM (ensemble of polyharmonic extreme learning machine) is trained on-line by the pixels sampling from the fixation and nonfixation area. Then the model of EPELM could be used to classify image pixels to form new binary fixation area. Depending upon the updated fixation area, the procedure of "pixel sampling-learning-classification" could be performed iteratively. If the previous binary fixation area and the latter one were similar enough in iteration, it indicates that the perception is saturated and the loop should be terminated. The binary output in iteration could be regarded as a kind of visual stimulation. So the multiple outputs of visual stimuli can be accumulated to form a new saliency map. Experiments on three image databases show the validity of our method. It can segment nucleated cells successfully in different imaging conditions.